xref: /openbmc/linux/include/uapi/linux/io_uring.h (revision ee65728e103bb7dd99d8604bf6c7aa89c7d7e446)
1 /* SPDX-License-Identifier: (GPL-2.0 WITH Linux-syscall-note) OR MIT */
2 /*
3  * Header file for the io_uring interface.
4  *
5  * Copyright (C) 2019 Jens Axboe
6  * Copyright (C) 2019 Christoph Hellwig
7  */
8 #ifndef LINUX_IO_URING_H
9 #define LINUX_IO_URING_H
10 
11 #include <linux/fs.h>
12 #include <linux/types.h>
13 
14 /*
15  * IO submission data structure (Submission Queue Entry)
16  */
17 struct io_uring_sqe {
18 	__u8	opcode;		/* type of operation for this sqe */
19 	__u8	flags;		/* IOSQE_ flags */
20 	__u16	ioprio;		/* ioprio for the request */
21 	__s32	fd;		/* file descriptor to do IO on */
22 	union {
23 		__u64	off;	/* offset into file */
24 		__u64	addr2;
25 		__u32	cmd_op;
26 	};
27 	union {
28 		__u64	addr;	/* pointer to buffer or iovecs */
29 		__u64	splice_off_in;
30 	};
31 	__u32	len;		/* buffer size or number of iovecs */
32 	union {
33 		__kernel_rwf_t	rw_flags;
34 		__u32		fsync_flags;
35 		__u16		poll_events;	/* compatibility */
36 		__u32		poll32_events;	/* word-reversed for BE */
37 		__u32		sync_range_flags;
38 		__u32		msg_flags;
39 		__u32		timeout_flags;
40 		__u32		accept_flags;
41 		__u32		cancel_flags;
42 		__u32		open_flags;
43 		__u32		statx_flags;
44 		__u32		fadvise_advice;
45 		__u32		splice_flags;
46 		__u32		rename_flags;
47 		__u32		unlink_flags;
48 		__u32		hardlink_flags;
49 		__u32		xattr_flags;
50 	};
51 	__u64	user_data;	/* data to be passed back at completion time */
52 	/* pack this to avoid bogus arm OABI complaints */
53 	union {
54 		/* index into fixed buffers, if used */
55 		__u16	buf_index;
56 		/* for grouped buffer selection */
57 		__u16	buf_group;
58 	} __attribute__((packed));
59 	/* personality to use, if used */
60 	__u16	personality;
61 	union {
62 		__s32	splice_fd_in;
63 		__u32	file_index;
64 	};
65 	union {
66 		struct {
67 			__u64	addr3;
68 			__u64	__pad2[1];
69 		};
70 		/*
71 		 * If the ring is initialized with IORING_SETUP_SQE128, then
72 		 * this field is used for 80 bytes of arbitrary command data
73 		 */
74 		__u8	cmd[0];
75 	};
76 };
77 
78 /*
79  * If sqe->file_index is set to this for opcodes that instantiate a new
80  * direct descriptor (like openat/openat2/accept), then io_uring will allocate
81  * an available direct descriptor instead of having the application pass one
82  * in. The picked direct descriptor will be returned in cqe->res, or -ENFILE
83  * if the space is full.
84  */
85 #define IORING_FILE_INDEX_ALLOC		(~0U)
86 
87 enum {
88 	IOSQE_FIXED_FILE_BIT,
89 	IOSQE_IO_DRAIN_BIT,
90 	IOSQE_IO_LINK_BIT,
91 	IOSQE_IO_HARDLINK_BIT,
92 	IOSQE_ASYNC_BIT,
93 	IOSQE_BUFFER_SELECT_BIT,
94 	IOSQE_CQE_SKIP_SUCCESS_BIT,
95 };
96 
97 /*
98  * sqe->flags
99  */
100 /* use fixed fileset */
101 #define IOSQE_FIXED_FILE	(1U << IOSQE_FIXED_FILE_BIT)
102 /* issue after inflight IO */
103 #define IOSQE_IO_DRAIN		(1U << IOSQE_IO_DRAIN_BIT)
104 /* links next sqe */
105 #define IOSQE_IO_LINK		(1U << IOSQE_IO_LINK_BIT)
106 /* like LINK, but stronger */
107 #define IOSQE_IO_HARDLINK	(1U << IOSQE_IO_HARDLINK_BIT)
108 /* always go async */
109 #define IOSQE_ASYNC		(1U << IOSQE_ASYNC_BIT)
110 /* select buffer from sqe->buf_group */
111 #define IOSQE_BUFFER_SELECT	(1U << IOSQE_BUFFER_SELECT_BIT)
112 /* don't post CQE if request succeeded */
113 #define IOSQE_CQE_SKIP_SUCCESS	(1U << IOSQE_CQE_SKIP_SUCCESS_BIT)
114 
115 /*
116  * io_uring_setup() flags
117  */
118 #define IORING_SETUP_IOPOLL	(1U << 0)	/* io_context is polled */
119 #define IORING_SETUP_SQPOLL	(1U << 1)	/* SQ poll thread */
120 #define IORING_SETUP_SQ_AFF	(1U << 2)	/* sq_thread_cpu is valid */
121 #define IORING_SETUP_CQSIZE	(1U << 3)	/* app defines CQ size */
122 #define IORING_SETUP_CLAMP	(1U << 4)	/* clamp SQ/CQ ring sizes */
123 #define IORING_SETUP_ATTACH_WQ	(1U << 5)	/* attach to existing wq */
124 #define IORING_SETUP_R_DISABLED	(1U << 6)	/* start with ring disabled */
125 #define IORING_SETUP_SUBMIT_ALL	(1U << 7)	/* continue submit on error */
126 /*
127  * Cooperative task running. When requests complete, they often require
128  * forcing the submitter to transition to the kernel to complete. If this
129  * flag is set, work will be done when the task transitions anyway, rather
130  * than force an inter-processor interrupt reschedule. This avoids interrupting
131  * a task running in userspace, and saves an IPI.
132  */
133 #define IORING_SETUP_COOP_TASKRUN	(1U << 8)
134 /*
135  * If COOP_TASKRUN is set, get notified if task work is available for
136  * running and a kernel transition would be needed to run it. This sets
137  * IORING_SQ_TASKRUN in the sq ring flags. Not valid with COOP_TASKRUN.
138  */
139 #define IORING_SETUP_TASKRUN_FLAG	(1U << 9)
140 
141 #define IORING_SETUP_SQE128		(1U << 10) /* SQEs are 128 byte */
142 #define IORING_SETUP_CQE32		(1U << 11) /* CQEs are 32 byte */
143 
144 enum io_uring_op {
145 	IORING_OP_NOP,
146 	IORING_OP_READV,
147 	IORING_OP_WRITEV,
148 	IORING_OP_FSYNC,
149 	IORING_OP_READ_FIXED,
150 	IORING_OP_WRITE_FIXED,
151 	IORING_OP_POLL_ADD,
152 	IORING_OP_POLL_REMOVE,
153 	IORING_OP_SYNC_FILE_RANGE,
154 	IORING_OP_SENDMSG,
155 	IORING_OP_RECVMSG,
156 	IORING_OP_TIMEOUT,
157 	IORING_OP_TIMEOUT_REMOVE,
158 	IORING_OP_ACCEPT,
159 	IORING_OP_ASYNC_CANCEL,
160 	IORING_OP_LINK_TIMEOUT,
161 	IORING_OP_CONNECT,
162 	IORING_OP_FALLOCATE,
163 	IORING_OP_OPENAT,
164 	IORING_OP_CLOSE,
165 	IORING_OP_FILES_UPDATE,
166 	IORING_OP_STATX,
167 	IORING_OP_READ,
168 	IORING_OP_WRITE,
169 	IORING_OP_FADVISE,
170 	IORING_OP_MADVISE,
171 	IORING_OP_SEND,
172 	IORING_OP_RECV,
173 	IORING_OP_OPENAT2,
174 	IORING_OP_EPOLL_CTL,
175 	IORING_OP_SPLICE,
176 	IORING_OP_PROVIDE_BUFFERS,
177 	IORING_OP_REMOVE_BUFFERS,
178 	IORING_OP_TEE,
179 	IORING_OP_SHUTDOWN,
180 	IORING_OP_RENAMEAT,
181 	IORING_OP_UNLINKAT,
182 	IORING_OP_MKDIRAT,
183 	IORING_OP_SYMLINKAT,
184 	IORING_OP_LINKAT,
185 	IORING_OP_MSG_RING,
186 	IORING_OP_FSETXATTR,
187 	IORING_OP_SETXATTR,
188 	IORING_OP_FGETXATTR,
189 	IORING_OP_GETXATTR,
190 	IORING_OP_SOCKET,
191 	IORING_OP_URING_CMD,
192 
193 	/* this goes last, obviously */
194 	IORING_OP_LAST,
195 };
196 
197 /*
198  * sqe->fsync_flags
199  */
200 #define IORING_FSYNC_DATASYNC	(1U << 0)
201 
202 /*
203  * sqe->timeout_flags
204  */
205 #define IORING_TIMEOUT_ABS		(1U << 0)
206 #define IORING_TIMEOUT_UPDATE		(1U << 1)
207 #define IORING_TIMEOUT_BOOTTIME		(1U << 2)
208 #define IORING_TIMEOUT_REALTIME		(1U << 3)
209 #define IORING_LINK_TIMEOUT_UPDATE	(1U << 4)
210 #define IORING_TIMEOUT_ETIME_SUCCESS	(1U << 5)
211 #define IORING_TIMEOUT_CLOCK_MASK	(IORING_TIMEOUT_BOOTTIME | IORING_TIMEOUT_REALTIME)
212 #define IORING_TIMEOUT_UPDATE_MASK	(IORING_TIMEOUT_UPDATE | IORING_LINK_TIMEOUT_UPDATE)
213 /*
214  * sqe->splice_flags
215  * extends splice(2) flags
216  */
217 #define SPLICE_F_FD_IN_FIXED	(1U << 31) /* the last bit of __u32 */
218 
219 /*
220  * POLL_ADD flags. Note that since sqe->poll_events is the flag space, the
221  * command flags for POLL_ADD are stored in sqe->len.
222  *
223  * IORING_POLL_ADD_MULTI	Multishot poll. Sets IORING_CQE_F_MORE if
224  *				the poll handler will continue to report
225  *				CQEs on behalf of the same SQE.
226  *
227  * IORING_POLL_UPDATE		Update existing poll request, matching
228  *				sqe->addr as the old user_data field.
229  */
230 #define IORING_POLL_ADD_MULTI	(1U << 0)
231 #define IORING_POLL_UPDATE_EVENTS	(1U << 1)
232 #define IORING_POLL_UPDATE_USER_DATA	(1U << 2)
233 
234 /*
235  * ASYNC_CANCEL flags.
236  *
237  * IORING_ASYNC_CANCEL_ALL	Cancel all requests that match the given key
238  * IORING_ASYNC_CANCEL_FD	Key off 'fd' for cancelation rather than the
239  *				request 'user_data'
240  * IORING_ASYNC_CANCEL_ANY	Match any request
241  */
242 #define IORING_ASYNC_CANCEL_ALL	(1U << 0)
243 #define IORING_ASYNC_CANCEL_FD	(1U << 1)
244 #define IORING_ASYNC_CANCEL_ANY	(1U << 2)
245 
246 /*
247  * send/sendmsg and recv/recvmsg flags (sqe->addr2)
248  *
249  * IORING_RECVSEND_POLL_FIRST	If set, instead of first attempting to send
250  *				or receive and arm poll if that yields an
251  *				-EAGAIN result, arm poll upfront and skip
252  *				the initial transfer attempt.
253  */
254 #define IORING_RECVSEND_POLL_FIRST	(1U << 0)
255 
256 /*
257  * accept flags stored in sqe->ioprio
258  */
259 #define IORING_ACCEPT_MULTISHOT	(1U << 0)
260 
261 /*
262  * IO completion data structure (Completion Queue Entry)
263  */
264 struct io_uring_cqe {
265 	__u64	user_data;	/* sqe->data submission passed back */
266 	__s32	res;		/* result code for this event */
267 	__u32	flags;
268 
269 	/*
270 	 * If the ring is initialized with IORING_SETUP_CQE32, then this field
271 	 * contains 16-bytes of padding, doubling the size of the CQE.
272 	 */
273 	__u64 big_cqe[];
274 };
275 
276 /*
277  * cqe->flags
278  *
279  * IORING_CQE_F_BUFFER	If set, the upper 16 bits are the buffer ID
280  * IORING_CQE_F_MORE	If set, parent SQE will generate more CQE entries
281  * IORING_CQE_F_SOCK_NONEMPTY	If set, more data to read after socket recv
282  */
283 #define IORING_CQE_F_BUFFER		(1U << 0)
284 #define IORING_CQE_F_MORE		(1U << 1)
285 #define IORING_CQE_F_SOCK_NONEMPTY	(1U << 2)
286 
287 enum {
288 	IORING_CQE_BUFFER_SHIFT		= 16,
289 };
290 
291 /*
292  * Magic offsets for the application to mmap the data it needs
293  */
294 #define IORING_OFF_SQ_RING		0ULL
295 #define IORING_OFF_CQ_RING		0x8000000ULL
296 #define IORING_OFF_SQES			0x10000000ULL
297 
298 /*
299  * Filled with the offset for mmap(2)
300  */
301 struct io_sqring_offsets {
302 	__u32 head;
303 	__u32 tail;
304 	__u32 ring_mask;
305 	__u32 ring_entries;
306 	__u32 flags;
307 	__u32 dropped;
308 	__u32 array;
309 	__u32 resv1;
310 	__u64 resv2;
311 };
312 
313 /*
314  * sq_ring->flags
315  */
316 #define IORING_SQ_NEED_WAKEUP	(1U << 0) /* needs io_uring_enter wakeup */
317 #define IORING_SQ_CQ_OVERFLOW	(1U << 1) /* CQ ring is overflown */
318 #define IORING_SQ_TASKRUN	(1U << 2) /* task should enter the kernel */
319 
320 struct io_cqring_offsets {
321 	__u32 head;
322 	__u32 tail;
323 	__u32 ring_mask;
324 	__u32 ring_entries;
325 	__u32 overflow;
326 	__u32 cqes;
327 	__u32 flags;
328 	__u32 resv1;
329 	__u64 resv2;
330 };
331 
332 /*
333  * cq_ring->flags
334  */
335 
336 /* disable eventfd notifications */
337 #define IORING_CQ_EVENTFD_DISABLED	(1U << 0)
338 
339 /*
340  * io_uring_enter(2) flags
341  */
342 #define IORING_ENTER_GETEVENTS		(1U << 0)
343 #define IORING_ENTER_SQ_WAKEUP		(1U << 1)
344 #define IORING_ENTER_SQ_WAIT		(1U << 2)
345 #define IORING_ENTER_EXT_ARG		(1U << 3)
346 #define IORING_ENTER_REGISTERED_RING	(1U << 4)
347 
348 /*
349  * Passed in for io_uring_setup(2). Copied back with updated info on success
350  */
351 struct io_uring_params {
352 	__u32 sq_entries;
353 	__u32 cq_entries;
354 	__u32 flags;
355 	__u32 sq_thread_cpu;
356 	__u32 sq_thread_idle;
357 	__u32 features;
358 	__u32 wq_fd;
359 	__u32 resv[3];
360 	struct io_sqring_offsets sq_off;
361 	struct io_cqring_offsets cq_off;
362 };
363 
364 /*
365  * io_uring_params->features flags
366  */
367 #define IORING_FEAT_SINGLE_MMAP		(1U << 0)
368 #define IORING_FEAT_NODROP		(1U << 1)
369 #define IORING_FEAT_SUBMIT_STABLE	(1U << 2)
370 #define IORING_FEAT_RW_CUR_POS		(1U << 3)
371 #define IORING_FEAT_CUR_PERSONALITY	(1U << 4)
372 #define IORING_FEAT_FAST_POLL		(1U << 5)
373 #define IORING_FEAT_POLL_32BITS 	(1U << 6)
374 #define IORING_FEAT_SQPOLL_NONFIXED	(1U << 7)
375 #define IORING_FEAT_EXT_ARG		(1U << 8)
376 #define IORING_FEAT_NATIVE_WORKERS	(1U << 9)
377 #define IORING_FEAT_RSRC_TAGS		(1U << 10)
378 #define IORING_FEAT_CQE_SKIP		(1U << 11)
379 #define IORING_FEAT_LINKED_FILE		(1U << 12)
380 
381 /*
382  * io_uring_register(2) opcodes and arguments
383  */
384 enum {
385 	IORING_REGISTER_BUFFERS			= 0,
386 	IORING_UNREGISTER_BUFFERS		= 1,
387 	IORING_REGISTER_FILES			= 2,
388 	IORING_UNREGISTER_FILES			= 3,
389 	IORING_REGISTER_EVENTFD			= 4,
390 	IORING_UNREGISTER_EVENTFD		= 5,
391 	IORING_REGISTER_FILES_UPDATE		= 6,
392 	IORING_REGISTER_EVENTFD_ASYNC		= 7,
393 	IORING_REGISTER_PROBE			= 8,
394 	IORING_REGISTER_PERSONALITY		= 9,
395 	IORING_UNREGISTER_PERSONALITY		= 10,
396 	IORING_REGISTER_RESTRICTIONS		= 11,
397 	IORING_REGISTER_ENABLE_RINGS		= 12,
398 
399 	/* extended with tagging */
400 	IORING_REGISTER_FILES2			= 13,
401 	IORING_REGISTER_FILES_UPDATE2		= 14,
402 	IORING_REGISTER_BUFFERS2		= 15,
403 	IORING_REGISTER_BUFFERS_UPDATE		= 16,
404 
405 	/* set/clear io-wq thread affinities */
406 	IORING_REGISTER_IOWQ_AFF		= 17,
407 	IORING_UNREGISTER_IOWQ_AFF		= 18,
408 
409 	/* set/get max number of io-wq workers */
410 	IORING_REGISTER_IOWQ_MAX_WORKERS	= 19,
411 
412 	/* register/unregister io_uring fd with the ring */
413 	IORING_REGISTER_RING_FDS		= 20,
414 	IORING_UNREGISTER_RING_FDS		= 21,
415 
416 	/* register ring based provide buffer group */
417 	IORING_REGISTER_PBUF_RING		= 22,
418 	IORING_UNREGISTER_PBUF_RING		= 23,
419 
420 	/* this goes last */
421 	IORING_REGISTER_LAST
422 };
423 
424 /* io-wq worker categories */
425 enum {
426 	IO_WQ_BOUND,
427 	IO_WQ_UNBOUND,
428 };
429 
430 /* deprecated, see struct io_uring_rsrc_update */
431 struct io_uring_files_update {
432 	__u32 offset;
433 	__u32 resv;
434 	__aligned_u64 /* __s32 * */ fds;
435 };
436 
437 /*
438  * Register a fully sparse file space, rather than pass in an array of all
439  * -1 file descriptors.
440  */
441 #define IORING_RSRC_REGISTER_SPARSE	(1U << 0)
442 
443 struct io_uring_rsrc_register {
444 	__u32 nr;
445 	__u32 flags;
446 	__u64 resv2;
447 	__aligned_u64 data;
448 	__aligned_u64 tags;
449 };
450 
451 struct io_uring_rsrc_update {
452 	__u32 offset;
453 	__u32 resv;
454 	__aligned_u64 data;
455 };
456 
457 struct io_uring_rsrc_update2 {
458 	__u32 offset;
459 	__u32 resv;
460 	__aligned_u64 data;
461 	__aligned_u64 tags;
462 	__u32 nr;
463 	__u32 resv2;
464 };
465 
466 /* Skip updating fd indexes set to this value in the fd table */
467 #define IORING_REGISTER_FILES_SKIP	(-2)
468 
469 #define IO_URING_OP_SUPPORTED	(1U << 0)
470 
471 struct io_uring_probe_op {
472 	__u8 op;
473 	__u8 resv;
474 	__u16 flags;	/* IO_URING_OP_* flags */
475 	__u32 resv2;
476 };
477 
478 struct io_uring_probe {
479 	__u8 last_op;	/* last opcode supported */
480 	__u8 ops_len;	/* length of ops[] array below */
481 	__u16 resv;
482 	__u32 resv2[3];
483 	struct io_uring_probe_op ops[0];
484 };
485 
486 struct io_uring_restriction {
487 	__u16 opcode;
488 	union {
489 		__u8 register_op; /* IORING_RESTRICTION_REGISTER_OP */
490 		__u8 sqe_op;      /* IORING_RESTRICTION_SQE_OP */
491 		__u8 sqe_flags;   /* IORING_RESTRICTION_SQE_FLAGS_* */
492 	};
493 	__u8 resv;
494 	__u32 resv2[3];
495 };
496 
497 struct io_uring_buf {
498 	__u64	addr;
499 	__u32	len;
500 	__u16	bid;
501 	__u16	resv;
502 };
503 
504 struct io_uring_buf_ring {
505 	union {
506 		/*
507 		 * To avoid spilling into more pages than we need to, the
508 		 * ring tail is overlaid with the io_uring_buf->resv field.
509 		 */
510 		struct {
511 			__u64	resv1;
512 			__u32	resv2;
513 			__u16	resv3;
514 			__u16	tail;
515 		};
516 		struct io_uring_buf	bufs[0];
517 	};
518 };
519 
520 /* argument for IORING_(UN)REGISTER_PBUF_RING */
521 struct io_uring_buf_reg {
522 	__u64	ring_addr;
523 	__u32	ring_entries;
524 	__u16	bgid;
525 	__u16	pad;
526 	__u64	resv[3];
527 };
528 
529 /*
530  * io_uring_restriction->opcode values
531  */
532 enum {
533 	/* Allow an io_uring_register(2) opcode */
534 	IORING_RESTRICTION_REGISTER_OP		= 0,
535 
536 	/* Allow an sqe opcode */
537 	IORING_RESTRICTION_SQE_OP		= 1,
538 
539 	/* Allow sqe flags */
540 	IORING_RESTRICTION_SQE_FLAGS_ALLOWED	= 2,
541 
542 	/* Require sqe flags (these flags must be set on each submission) */
543 	IORING_RESTRICTION_SQE_FLAGS_REQUIRED	= 3,
544 
545 	IORING_RESTRICTION_LAST
546 };
547 
548 struct io_uring_getevents_arg {
549 	__u64	sigmask;
550 	__u32	sigmask_sz;
551 	__u32	pad;
552 	__u64	ts;
553 };
554 
555 #endif
556